Fix: Could not load dynamic library 'libcudart.so.11.0'

Could not load dynamic library 'libcudart.so.11.0'; dlerror: libcudart.so.11.0: cannot open shared object file means TensorFlow went looking for the CUDA runtime it was built against and the dynamic loader couldn't find it. libcudart is the CUDA Runtime library, and the 11.0 tells you exactly which one it needs.

What "11.0" really requires

The 11.0 is a SONAME — a version tag the binary links against. Two facts make it manageable:

• It's a major-version tag. From CUDA 11 onward, NVIDIA guarantees minor-version compatibility, so every CUDA 11.x release ships the same libcudart.so.11.0. You do not need exactly 11.0 — 11.2, 11.6, 11.8 all provide it, as long as your NVIDIA driver is ≥ 450.
• CUDA 12 does not count. CUDA 12.x ships libcudart.so.12, a different file. Having "CUDA installed" isn't enough if it's the wrong major version.

One more thing worth knowing: on TensorFlow this is frequently a non-fatal warning. TensorFlow logs it, disables the GPU, and continues on CPU. So first, decide whether it even hurts you:

import tensorflow as tf
print(tf.config.list_physical_devices('GPU'))

An empty list means the missing runtime cost you GPU acceleration — worth fixing. A listed GPU means you can ignore the message.

Triage in one command

ldconfig -p | grep libcudart

• It prints a path → the runtime exists; this is a path problem. The loader just doesn't know where it is.
• It prints nothing → the CUDA 11.x runtime isn't installed. This is a missing-runtime problem.

Confirm what you have with nvcc --version (toolkit) and nvidia-smi (driver + its CUDA ceiling).

The fixes

Path problem — expose the existing runtime:

export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:$LD_LIBRARY_PATH
sudo ldconfig

Missing runtime — install a matching CUDA, the easy way:

# modern TensorFlow can bring its own matched CUDA:
pip install "tensorflow[and-cuda]"
 
# or let conda resolve a compatible runtime:
conda install -c conda-forge cudatoolkit=11.8
 
# or run the official GPU image (only the host driver is needed):
docker run --gpus all -it tensorflow/tensorflow:2.14.0-gpu bash

Or move TensorFlow to your CUDA: if you're pinned to CUDA 12, install a TensorFlow build compiled for it instead of downgrading CUDA. Match either direction — the framework's expected CUDA major and the installed CUDA major must be the same.

Related

This is one member of the CUDA shared-library-load family (libcublas, libcudnn, libcufft, libnvinfer, …) — all the same SONAME-matching problem. For the full decision tree covering every variant, see the CUDA and cuDNN library load error diagnostic guide.

Prevent it

• Match the major version deliberately using TensorFlow's CUDA/cuDNN compatibility table before installing.
• Prefer pip install tensorflow[and-cuda], conda, or Docker so the runtime is resolved for you.
• Keep driver and toolkit straight: CUDA 11.x needs driver ≥ 450; the driver sets the ceiling, the toolkit provides the runtime TensorFlow links.

Decide whether it's a path or a missing-runtime problem with one ldconfig, then either expose the existing libcudart.so.11.0 or install any CUDA 11.x that provides it.

Sources

• TensorFlow API docs